Introduction
In scientific machine learning, the pursuit of state-of-the-art performance has traditionally been a one-dimensional race toward higher numerical accuracy. While a 0.1% improvement in relative error is often celebrated, the environmental cost—thousands of GPU hours and significant carbon emissions—is rarely reported. As neural partial differential equation (PDE) solvers scale from toy problems to complex aerospace and climate simulations, their carbon footprint is skyrocketing. Evaluating models solely on accuracy creates a bias toward computationally wasteful architectures. To address this, we introduce EcoL2, a holistic evaluation metric that balances numerical precision with the total carbon footprint of a model's lifecycle.
The EcoL2 Metric: A Holistic Framework
The EcoL2 metric integrates accuracy and sustainability into a single score between 0 and 1. It consists of an accuracy-based numerator and a carbon-based denominator. The numerator uses an exponential-log transformation of the relative L2 error ($R$), ensuring that gains in high-accuracy regimes are weighted more heavily. The denominator aggregates the total lifecycle carbon emissions ($C$), scaled by a hyperparameter $\beta$.
The "Total Lifecycle Carbon" is defined as the sum of four distinct stages: embodied carbon from data generation, developmental carbon from hyperparameter tuning, operational carbon from training, and inference carbon from deployment. By adjusting hyperparameters $\alpha$ and $\beta$, the metric can be tailored to prioritize either extreme accuracy for high-stakes engineering or extreme efficiency for large-scale deployment.
Experimental Results and Insights
The authors evaluated EcoL2 across several canonical PDE benchmarks, including Advection, Reaction, Wave, Korteweg–de Vries (KdV), and Kuramoto–Sivashinsky (KS) equations. These tests involved various architectures such as Physics-Informed Neural Networks (PINNs) and Neural Operators (DON, FNO, CNO).
Accuracy is Deceptive
A key finding is that models with nearly identical accuracy can have vastly different environmental impacts. For the Advection equation, PINNsFormer and SPINN achieved similar relative errors (~4e-4). However, PINNsFormer’s transformer-based architecture resulted in a carbon footprint nearly five times larger than SPINN's, leading to a significantly lower EcoL2 score.
The "Tuning Tax" and Geography
Developmental carbon ($C_d$) often emerges as the dominant contributor to the total footprint. The "hidden" cost of finding optimal hyperparameters can emit significantly more CO2 than the final training run itself. Furthermore, the geographical location of the compute server plays a critical role. Due to differences in grid carbon intensity, an FNO model trained in a coal-heavy region can produce an order of magnitude more CO2 than the same model trained in a renewables-heavy region, even with identical compute time.
Conclusion
The EcoL2 metric provides a necessary shift in how we benchmark scientific machine learning models. By ranking models based on "bang for the buck," it identifies winners that are both performant and sustainable. While challenges remain in accurately attributing embodied carbon for shared datasets and relying on third-party emission trackers, the adoption of carbon accounting is a vital step toward responsible AI. Future research should treat carbon reporting as a standard requirement, alongside accuracy and runtime, to incentivize the development of greener, more efficient neural solvers.
Summary updated